Dietary patterns and changes in frailty status: the Rotterdam study

This study was embedded in the Rotterdam Study (RS)–an ongoing prospective cohort in the Netherlands [19]. A more detailed description of the RS is provided elsewhere [19]. Briefly, the first baseline visits took place between 1990 and 1993. All residents aged 55 years and over in the Ommoord district of Rotterdam (n = 10,215), the Netherlands, were invited to participate, of which 7983 (78%) took part in the RS’s first cohort (RS-I). The study was extended in the year 2000 (RS-II; n = 3011) and in 2006, inviting all residents aged 45 years and over (RS-III; n = 3932). In total, 14,926 participants were included in the RS, who visited the research center for detailed measurements every 3–4 years. During an extensive home interview, trained research assistants collected data on a broad range of health variables including, activities of daily living, current health status, use of medication, depression and lifestyle. Subsequently, participants visited the study center for detailed examinations with an emphasis on imaging, collection of body fluids, and physical functioning. The RS was approved by the Medical Ethics Committee of the Erasmus Medical Center and by the review board of The Netherlands Ministry of Health, Welfare and Sports. All participants signed an informed consent. This study adheres to the Declaration of Helsinki for research involving human subjects.

For the current study, we included the first and second visit of the third cohort of the RS (RS-III-1and RS-III-2) comprising of 3932 participants. For 2632 participants, valid dietary intake and a frailty index were available at baseline (2006–2008) and for 2253 participants a frailty index was also available at follow-up (2012–2013, Fig. 1).

Dietary intake was measured with a self-administrated semi-quantitative food frequency questionnaire (FFQ) developed by Wageningen University and Research centre, adapted for the Rotterdam Study. The ability of the FFQ to rank people according to their intake was previously shown in two validation studies using a 9-day dietary record [20] and a 4-week dietary history [21]. The FFQ includes 389 items about the frequency and amount of consumed food items in days, weeks and months according to the previous year and was filled out at home. For the estimation of the portion sizes in grams standardized household measures were applied [22]. For calculation of the nutritional data the Dutch Food Composition Table (NEVO) of 2006 was used [23]. Participants with extremely high (>5000 kcal) or low (<500 kcal) daily energy intake were excluded as it was assumed that their questionnaire was unreliable (Fig. 1).

Two different approaches to determine dietary patterns were applied: (1) an a priori defined index for diet quality and (2) a posteriori defined dietary patterns using principal component analysis (PCA).

Frailty was measured with a frailty index, an instrument based on the accumulation of health deficits [2]. In general, deficits can be symptoms, signs, diseases, disabilities and laboratory measurements as long as they are age-related and health-related and are not too exceptional or too common [3]. We used a slightly adapted version of a previous validated frailty index designed for the Rotterdam Study, consisting of 38 health-related variables covering several health domains: functional status (n = 13), health conditions (n = 6), cognition (n = 6), diseases (n = 6), nutritional status (n = 3) and mood (n = 4) [28]. Deficits were dichotomized or categorized into a score ranging from 0 (not present) till 1 (present). Per person, the number of present deficits was divided by the total number of deficits, providing a continuous score ranging from 0 (no deficits present, least frail) till 1 (all deficits present, extremely frail). Missing values on the deficits were imputed using multiple imputation by chained equations [28]. Individuals with less than 20 observed items were determined to have insufficient information to considerably contribute to a valid frailty index and were excluded from the analyses (Fig. 1). To be able to evaluate changes over time we had to remove seven items from the original Rotterdam Study Frailty Index, namely: vitamin D, sex hormone binding globulin, mobility, uric acid, proBNP, CRP and homocysteine. Because, unfortunately, these biomarkers were not assessed at follow-up. Characteristics of the original Rotterdam Study frailty index and the adapted version are provided in Supplementary Table III, no major differences in the mean or median were observed. Furthermore, the two scales had a high mutual correlation (r = 0.98) and similar associations with age and mortality (Supplementary Table III).

Height (cm) and body weight (kg) were measured at the research center using a stadiometer wearing light clothing. Body mass index (BMI) was calculated as body weight (kg) divided by height (m)². Smoking status was classified as never, former or current smoker. Level of education was determined by the highest attained education and classified as low (primary education and lower vocational education), middle (secondary general education and secondary vocational education), middle-high (higher general education) or high (higher vocational education or university education). Monthly household income was classified as low (<€1.500), middle (€1.500–2.900) or high (>€2.900). Physical activity was assessed with the LASA physical activity Questionnaire (LAPAQ) and metabolic equivalents (MET) scores were calculated as the sum of hours a week spent in light, moderate or vigorous activity (walking, cycling, gardening, sports, and hobbies), expressed in metabolic equivalent of task (MET) score [29]. MET scores represent the energy that is required for an activity divided by the energy necessary at rest [30]. Total energy intake in kilocalories per day and use of dietary supplements (yes/no) were all retrieved from the FFQ.

First, baseline characteristics of the study population were shown in strata of frailty, dichotomized at the median. A p value for the observed values was provided using independent sample t tests for continuous variables and X ² for categorical variables. Second, linear regression analyses were performed to examine the cross-sectional associations between adherence to each dietary pattern and the frailty index at baseline (all in Z-scores). Analyses were performed as a basic model, adjusted for age and sex (model 1), followed by a model that was additionally adjusted for smoking, level of education and income, physical activity, dietary supplement use (model 2), and a model additionally adjusted for energy intake (model 3). Confounders were tested based on previous studies [31, 31] and included in the models if they substantial change in effect-estimate on at least one of the dietary patterns (>10%). Additionally, the three a posteriori derived dietary patterns were adjusted for each other. The third models were adjusted for total energy intake because by design of the DHD-index, it might be easier to adhere to the guidelines at higher levels of energy intake. Third, we evaluated the association between the dietary patterns and changes in frailty index over time. To test if the frailty index changed significantly over time we applied a paired t test. Thereafter, in line with the cross-sectional results we created a basic model and an adjusted model, using the frailty index at follow-up as an outcome, additionally adjusting for the baseline frailty index. The coefficients of this model can be interpreted as the difference between the mean change frailty index score for each unit increase in exposure [32]. For the a posteriori defined dietary patterns we calculated the food group intakes corresponding to 1SD difference in dietary pattern adherence to increase the interpretability of the results. To exclude the possibility that results were driven by nutrition-associated deficits in the frailty index, we created a frailty index without BMI, HDL and total cholesterol and reran the analyses.

Additionally, we performed several sensitivity analyses using the cross-sectional data. To test potential selection bias, we calculated and compared the frailty index score for participants included and excluded in the main analyses. We tested for potential interaction by adding the product term of adherence to each of the dietary patterns with total energy intake to model 3. A similar approach was used to study interaction with sex, age and BMI. Stratified analyses were performed if the p for interaction was <0.05. Last, we performed the analyses in subgroups after excluding (1) participants with incomplete dietary intake data (>1% missing items in the FFQ), and (2) participants who deceased within 3 years after baseline. Analyses were performed using SPSS statistical software (IBM, version 23). A p value <0.05 was considered statistically significant.

The median (interquartile range) frailty index of our population (n = 2632) was 0.14 (0.09, 0.19). Characteristics of our study population in strata of the frailty index above and below the median are shown in Table 1. On average, participants were 57 years (SD = 7.2) and the mean DHD-index for the full population was 56.2 (SD 9.28).

A posteriori, we derived three population-specific dietary patterns that we labeled: (1) a “Traditional” pattern, characterized by a high intake of savory snacks, legumes, eggs, fried potatoes, alcohol, processed meat and soup; (2) a “Carnivore” pattern, characterized by a high intake of red meat and poultry with a low intake of meat replacements; and (3) a “Health Conscious” pattern, characterized by a high intake of whole grains, vegetables, fruit and nuts. The factor loadings of the food groups are presented in Table 2. The “Traditional” pattern explained 10.0%, the “Carnivore” pattern 7.7% and the “Health Conscious” pattern 5.4% of the total variance in food group intake (Table 2). The DHD-index was positively associated with the “Traditional” pattern (Pearson’s r = 0.37) and with the “Health Conscious” pattern (Pearson’s r = 0.13), and negatively associated with the “Carnivore” pattern (Pearson’s r = −0.25).

In the fully adjusted models, a priori defined adherence to the DHD-index was associated with lower frailty index scores (β (95% CI) = −0.07 (−0.10, −0.03), Table 3, model 3. This implies that with every SD increase in DHD-index (1 SD = 9.28) the frailty index was on average 0.05 SD lower (1 SD = 0.08). More specifically, any of the following was associated with a 0.08 lower frailty index score: 200 g of vegetables, 14 g fibers per 100 kcal a day or <1 energy % trans fatty acids (supplementary data Table 1). After adjustment for all covariates and energy intake, of the a posteriori pattern only the “Carnivore” pattern was associated with frailty (Table 3). The interpretation of the Z-scores (one SD difference) for each dietary pattern is provided in Supplementary Table V.

In total, 2253 participants were included in the longitudinal analyses. For these participants, the median frailty index at follow-up was 0.14 (SD = 0.08) and frailty index was 0.007 lower at follow-up than at baseline (SD = 0.06, p value <0.001). Higher adherence to the a priori defined DHD-index and the a posteriori “Traditional” pattern at baseline were associated with reduced frailty indices overtime in all models, β (95% CI) = −0.07 (−0.10, −0.04) and β (95% CI) = −0.07 (−0.11, −0.04), respectively, Table 4, model 3. These results imply that with any of the following the frailty index at follow-up was 0.005 lower: 21 g/day increase in refined grain products, 13 g/day increase in potatoes or 18 g/day increase in savory snacks (supplemental table V). Adherence to the a posteriori derived “Carnivore” pattern was associated with an increased frailty index overtime in model 2, β (95% CI) = 0.03 (0.00, 0.07), but these results were no longer significant if adjusted for energy intake. The a posteriori defined “Health conscious” pattern was not associated with changes in frailty.

We observed that excluded participants had on average a higher frailty index score (mean 0.18) than the included participants (mean 0.14, p value <0.001). Additional adjustment by BMI did not highly influence the results. However, the cross-sectional analyses between adherence to the “Carnivore” pattern with frailty was no longer significant. We did not observe significant interaction terms between any of the dietary patterns and gender (p value range 0.13–0.86), total energy intake (p value range 0.06–0.60) or BMI (p value range 0.27–0.94) on frailty. Supplementary Table IV shows the sensitivity analyses for the cross-sectional results. Using the original 45-item frailty index did not influence the association between the a priori derived DHD-index and frailty (Supplementary Table IV). Excluding participants that deceased within 3 years (n = 38) or participants with incomplete FFQ data (n = 867) provided similar results for the a priori derived DHD-index, whereas the a posteriori derived “Carnivore” pattern was significantly associated with higher frailty scores when excluding participants who died within 3 years and participants without fully complete FFQs (Supplementary Table IV). Last, although effect estimates were similar, we observed a slightly stronger association between the DHDI index and frailty in participants aged above the median (57 years) than in those below the median age (Supplementary Table IV).